Data set screening method, device, equipment, medium and product

Abstract

The present disclosure relates to a data set screening method, device, equipment, medium and product, the method comprising: obtaining task portrait information of a candidate task, the task portrait information representing the priority of data in at least one quality dimension; determining conflict activation information of the candidate task according to the task portrait information and conflict information; and screening a candidate data set of the candidate task according to the conflict activation information to obtain a target data set. The present method explicitly determines the priority of each quality dimension through the task portrait, generates conflict activation information in combination with the conflict information, explicitly determines the dimensions that are necessary for the task and have structural contradictions in the data set, and finally screens the data in the candidate data set based on the conflict activation information, thereby realizing accurate screening and optimization of the candidate data set, effectively solving the problems of unclear data quality priority and potential conflicts, improving the accuracy and consistency of the target data set, and significantly improving the efficiency and quality of subsequent task processing.

Description

Technical Field

[0001] This disclosure relates to the field of cloud computing technology, and more specifically, to a method, apparatus, device, medium, and product for filtering datasets. Background Technology

[0002] As large-scale model training adopts the paradigm of "model and data co-optimization," dataset quality is replacing model architecture as the core variable determining the upper limit of model capabilities. Studies such as LIMA and Phi have shown that a small amount of carefully selected data can surpass coarsely labeled datasets of much larger magnitude during the SFT stage. The engineering evaluation and tuning of data quality has become a key link in the engineering process of large-scale models.

[0003] In related technologies, optimization methods use the same filtering method for different data, ignoring the differences between different data. Summary of the Invention

[0004] The purpose of this disclosure is to provide a method, apparatus, device, medium, and product for filtering datasets.

[0005] To achieve the above objectives, in a first aspect, this disclosure provides a method for filtering datasets, the method comprising: Obtain task profile information for candidate tasks, wherein the task profile information represents the priority of data in at least one quality dimension; Based on the task profile information and conflict information, determine the conflict activation information of the candidate task; Based on the conflict activation information, the candidate datasets for the candidate tasks are filtered to obtain the target dataset.

[0006] Optionally, the step of filtering the candidate datasets for the candidate tasks based on the conflict activation information to obtain the target dataset includes: Based on the conflict activation information, a first subset of data in the candidate dataset that has a conflict quality dimension and a second subset of data that does not have a conflict quality dimension are determined. Based on the quality dimension of the conflict, determine the first weight corresponding to each data item in the first data subset; Based on the weight of each quality dimension, determine the second weight corresponding to each data point in the second data subset; The data in the candidate dataset is filtered according to the first weight and the second weight to obtain the target dataset.

[0007] Optionally, determining the first weight corresponding to each data item in the first data subset based on the quality dimension of the conflict includes: Based on the quality dimension of the conflict, obtain the marginal contribution information of each data point in the first data subset under the quality dimension of each conflict, and determine the marginal contribution information as the first weight of each data point in the quality dimension of each conflict.

[0008] Optionally, the step of filtering the data in the candidate dataset according to the first weight and the second weight to obtain the target dataset includes: Based on the first weight and the second weight, determine the target weight for each data item in the candidate dataset; The data in the candidate dataset are sorted according to the target weights, and the target dataset is determined based on the sorted data.

[0009] Optionally, obtaining the task profile information of the candidate task includes: Obtain the task information of the candidate task, the task information including task type and task metadata; Based on the task information and quality dimensions, the task profile information of the candidate tasks is determined.

[0010] Optionally, the method further includes: Historical sample pairs were determined based on historical datasets, historical quality dimensions, and ablation experiment result sets; Obtain the conflict coefficient for each pair of historical samples for each pair of historical quality dimensions; The conflict information is determined based on the conflict coefficient.

[0011] Optionally, the method further includes: Based on the target dataset, determine the latest ablation experiment result set; The conflict information is updated based on the latest ablation experiment results set.

[0012] Secondly, this disclosure provides a dataset filtering apparatus, the apparatus comprising: The information acquisition module is configured to acquire task profile information of candidate tasks, wherein the task profile information represents the priority of data in at least one quality dimension. The data filtering module is configured to determine the conflict activation information of the candidate tasks based on the task profile information and conflict information. The data acquisition module is configured to filter the candidate datasets of the candidate tasks based on the conflict activation information to obtain the target dataset.

[0013] Thirdly, this disclosure provides an electronic device, including: A memory on which computer programs are stored; A processor for executing the computer program in the memory to implement the steps of the method of any one of the first aspects.

[0014] Fourthly, this disclosure provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method described in any of the first aspects.

[0015] Fifthly, this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in any one of the first aspects.

[0016] The above technical solution clarifies the priority of each quality dimension through task profiling, generates conflict activation information by combining conflict information, identifies the dimensions in the dataset that are necessary for the task and have structural contradictions, and finally filters the data in the candidate dataset based on the conflict activation information. This achieves accurate filtering and optimization of the candidate dataset, effectively solves the problems of unclear data quality priorities and potential conflicts, improves the accuracy and consistency of the target dataset, and significantly improves the efficiency and quality of subsequent task processing.

[0017] Other features and advantages of this disclosure will be described in detail in the following detailed description section. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the present disclosure and form part of the specification. They are used together with the following detailed description to explain the present disclosure, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a method for filtering a dataset according to an exemplary embodiment.

[0019] Figure 2 This is a flowchart illustrating a method for filtering a dataset according to another exemplary embodiment.

[0020] Figure 3 This is a block diagram illustrating a dataset filtering device according to an exemplary embodiment.

[0021] Figure 4 This is a block diagram illustrating an electronic device according to an exemplary embodiment. Detailed Implementation

[0022] The specific embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this disclosure.

[0023] It should be noted that all actions involving the acquisition of signals, information, or data in this disclosure are carried out in compliance with the relevant data protection laws and policies of the country where the location is situated, and with authorization from the owner of the relevant device.

[0024] To address the technical problems raised in the background, this application provides a method for filtering datasets. Figure 1 This is a flowchart illustrating a method for filtering a dataset according to an exemplary embodiment, the method comprising the following steps.

[0025] In step S11, task profile information of candidate tasks is obtained, wherein the task profile information represents the priority of data in at least one quality dimension.

[0026] Task profiling information refers to information used to characterize the priority of candidate task data across different quality dimensions, reflecting the rigid requirements and flexibility of each candidate task's data regarding each quality dimension. Quality dimensions include, but are not limited to, consistency, accuracy, timeliness, compliance, diversity, and contamination-free aspects. Different candidate tasks and different data have different priority requirements for quality dimensions. For example, for tasks training image recognition models, data diversity and accuracy have higher priority requirements; while for tasks training text generation models, data consistency and compliance have relatively higher priority requirements.

[0027] In step S12, the conflict activation information of the candidate task is determined based on the task profile information and the conflict information.

[0028] Conflict information refers to priority conflicts or resource competition among different quality dimensions. For example, high diversity requirements may reduce data consistency, while high accuracy requirements may sacrifice data timeliness. Conflict information can be determined based on a pre-defined conflict rule base or historical conflict records.

[0029] Furthermore, conflict activation information refers to the conflict status indicator triggered by conflicting priorities in candidate tasks under a specific combination of quality dimensions. Conflict activation information can be a conflict activation mask, used to mark which quality dimensions have conflicting relationships that require special handling. For example, when a candidate task simultaneously requires high diversity and high consistency, the bits corresponding to these two dimensions in the conflict activation mask will be set to 1, indicating that the task has a conflict under this dimension combination that needs to be addressed. Conflict activation information can also be represented in vector form, indicating the intensity or level of conflict between different quality dimensions of the data. By determining the conflict activation information, the combination of quality dimensions that should be prioritized or avoided during the data screening process can be determined, thus providing a basis for subsequent data screening strategies.

[0030] In step S13, the candidate datasets of the candidate tasks are filtered according to the conflict activation information to obtain the target dataset.

[0031] One method for determining the target dataset is to assign a quality score to each data sample in the candidate dataset and adjust the score weights based on conflict activation information to select high-quality data samples that meet the task requirements. Another method is to identify data with quality dimension conflicts and data without such conflicts based on conflict activation information. For data with quality dimension conflicts, different weights for each quality dimension can be used to determine its applicability score in the candidate task. For data without quality dimension conflicts, the original scores for each quality dimension can be used for comprehensive evaluation. In this way, the scored data can be sorted or thresholded to obtain the target dataset that meets the task requirements.

[0032] In the solution provided in this application embodiment, the priority of each quality dimension is clarified by task profiling, and then conflict activation information is generated by combining conflict information to identify the dimensions in the dataset that are necessary for the task and have structural contradictions. Finally, the data in the candidate dataset is filtered based on the conflict activation information, which realizes the accurate filtering and optimization of the candidate dataset, effectively solves the problems of unclear data quality priority and potential conflicts, improves the accuracy and consistency of the target dataset, and significantly improves the efficiency and quality of subsequent task processing.

[0033] Figure 2 This is a flowchart illustrating a method for filtering a dataset according to another exemplary embodiment, the method comprising the following steps.

[0034] In step S21, the task information of the candidate task is obtained, including the task type and task metadata.

[0035] The task types include, but are not limited to, pre-training, SFT, RLHF, evaluation sets, multimodal, and agent types. Task metadata includes, but is not limited to, information such as the target domain, target model size, and target metrics. For example, for image classification tasks, task metadata might include image resolution, number of categories, and annotation specifications; for text generation tasks, it might include language, text length range, and domain type.

[0036] In step S22, the task profile information of the candidate task is determined based on the task information and the quality dimension.

[0037] In this embodiment, task profile information can be determined based on the mapping relationship between task information and quality dimensions, calculated through preset rules or models. Alternatively, task profile information can be determined based on historical task data, establishing a correlation model between task information and quality dimensions, and outputting corresponding task profile information by inputting current task information. Through task profile information, the importance weight of different quality dimensions in a specific task can be clearly defined, thus providing more targeted guidance for subsequent data filtering.

[0038] In step S23, the conflict activation information of the candidate task is determined based on the task profile information and the conflict information.

[0039] For a detailed explanation of step S23, please refer to the detailed explanation of step S12, which will not be repeated here.

[0040] In some implementations, determining the conflict coefficient includes: identifying historical sample pairs based on historical datasets, historical quality dimensions, and ablation experiment result sets; obtaining the conflict coefficient for each historical sample pair for each historical quality dimension; and determining the conflict information based on the conflict coefficient.

[0041] The historical dataset refers to the dataset from historical tasks, which includes, but is not limited to, data samples used in those tasks. Historical quality dimensions refer to the various dimensions used to evaluate data quality in historical tasks, such as diversity, consistency, and accuracy. The ablation experiment result set records the changes in model performance after removing or adjusting specific quality dimensions in historical tasks, thus providing empirical evidence for determining the importance of each quality dimension.

[0042] Furthermore, historical sample pairs can be any two data points selected from the historical dataset, or they can be two data points that conflict in terms of historical quality. Historical sample pairs can also be selected based on the best or worst performing data combinations in historical tasks. No restrictions are placed on the method for determining historical sample pairs here.

[0043] Specifically, the conflict coefficient can be determined by calculating the performance differences of historical sample pairs across different historical quality dimensions, and combining this with the impact of each dimension on model performance in the ablation experiment results set, resulting in a quantitative index after comprehensive evaluation. Alternatively, the conflict coefficient can be determined by analyzing the degree of conflict between historical sample pairs on specific quality dimensions, combined with the dimension importance weights revealed by the ablation experiment, thereby determining a conflict coefficient that reflects the interaction between dimensions.

[0044] In this embodiment, the determination of conflict information can be based on the historical quality dimension or on the distribution pattern of the conflict coefficient in different tasks.

[0045] For example, in terms of historical quality dimension The historical dataset pool and ablation experiment result set are For example, among which, It is a historical dataset. It is its corresponding task. It is the downstream indicator change vector, which refers to the change in model performance indicators after removing or adjusting a specific quality dimension in the ablation experiment, reflecting the degree of influence of that dimension on the quality of task completion.

[0046] In the process of determining conflict information, each historical dataset is first processed. Calculate its score vectors for each dimension. And construct historical sample pairs. For each pair of historical quality dimensions Calculate the conflict coefficient The formula is as follows: (1) in, For the task The conflict coefficient between quality dimension i and quality dimension j, with a range of values. , A value greater than 0 indicates a conflict between the two quality dimensions. <0 indicates collaboration between the two quality dimensions. ≈0 indicates that there is no significant correlation between the two quality dimensions. For task identification, For the task Under these conditions, the change in downstream indicators The partial derivative of the score for the i-th quality dimension represents the marginal sensitivity of the i-th quality dimension to downstream task metrics. For the task Under these conditions, the change in downstream indicators Score for the j-th quality dimension The partial derivative of represents the marginal sensitivity of the j-th quality dimension to downstream task indicators. For the task The corresponding change in the evaluation index is the scalar result obtained from the ablation experiment. The score of the i-th quality dimension; Let be the score for the j-th quality dimension.

[0047] In some specific embodiments, it can also be based on historical tasks, historical sample pairs, and Established in different tasks The third-order tensor on This refers to the dimensional conflict map under task conditions. By analyzing the patterns of this tensor across different task slices, a dimensional conflict map can be established to identify both universal and task-specific combinations of conflict dimensions. It can also provide prior knowledge about conflict dimensions for subsequent tasks.

[0048] In some specific embodiments, the method further includes: determining the latest ablation experiment result set based on the target dataset; and updating the conflict information based on the latest ablation experiment result set.

[0049] In this embodiment, to ensure more accurate conflict information, ablation experiments are conducted based on the target dataset, and the conflict information is updated according to the latest ablation experiment results. Conflict information can be determined each time the target dataset is identified, or it can be determined when the number of times the target dataset is identified reaches a preset number. In this case, the conflict information can be updated based on all target datasets, or it can be updated based on one or more target datasets. Conflict information can also be updated based on the number of new samples received. When the number of new samples exceeds a preset number, the conflict information is updated; for example, when 50 new samples are identified, the conflict information is updated. Conflict information can also be updated based on a preset time period, such as weekly or monthly.

[0050] For example, conflict information can be updated based on the filtered target dataset. Candidate proxy model; Downstream evaluation metrics This is achieved by using a candidate proxy model, which is a model used to simulate the target dataset, and downstream evaluation metrics to measure the performance of the candidate proxy model on the target dataset.

[0051] use Fine-tune the agent model and calculate To determine the ablation experiment results for the target dataset.

[0052] like The prediction direction is significantly inconsistent with conflict information, such as the increased diversity of conflict information predictions, but... Characterizing the measured decrease, Add the results to the historical ablation experiment set H. Updates can be made in real-time based on new samples in the historical ablation experiment set H, or, to save computational resources, conflict information can be updated in batches after H has accumulated to a certain size. Specifically, when the number of samples in H reaches a preset threshold, or after a preset time interval, conflict information is retrained or updated based on all samples in H.

[0053] In step S24, based on the conflict activation information, a first subset of data in the candidate dataset that has a conflict quality dimension and a second subset of data that does not have a conflict quality dimension are determined.

[0054] In this embodiment, conflict dimension detection can be performed on each data point in the candidate dataset based on the conflict activation information. Specifically, for each data point, it can be input into a conflict detection model. If the activation value output by the model exceeds a preset threshold, the data is determined to have a conflict quality dimension and is assigned to the first data subset. Otherwise, it is assigned to the second data subset. This division effectively distinguishes data with potential conflicts. Alternatively, the first and second data subsets in the candidate dataset can be determined by calculating a task-specific conflict switch table (CAM matrix).

[0055] For example, let's take calculating the CAM matrix to determine the first and second subsets of data in the candidate dataset. The conflict information is... Task profile information .

[0056] First, determine the dimensional activation vectors based on the task profile information, and retain only the essential quality dimensions from the candidate data. .

[0057] The CAM matrix is ​​calculated based on the dimensional activation vector, as shown in the following formula: (2) in, Let be the matrix element in the i-th row and j-th column of the CAM matrix, representing the connection strength from the i-th dimension to the j-th dimension; For the task The conflict coefficient between the i-th quality dimension and the j-th quality dimension; As a task identifier; The activation flag for the i-th dimension takes a value of 0 or 1 and is obtained from the dimension activation vector α = ReLU(π(τ)). =1 indicates that the i-th dimension is a quality dimension required for the candidate task; similarly, This is the activation flag for the j-th dimension; For indicator functions, when When true, the value is 1. The value is 0 when the condition is not met. This is a preset conflict threshold. Only when the dimension... and dimensions These are all essential quality dimensions, and the dimensions and dimensions The CAM matrix is ​​activated only when the conflict coefficient for the candidate task is greater than a threshold. The conflict threshold can be 0.3, 0.5, or can be set according to the actual situation; no limit is set here.

[0058] The dimension pairs corresponding to the non-zero elements in the CAM matrix are defined as activation conflict sets. , .

[0059] based on The CAM matrix determines the first subset of data in the candidate dataset that has conflicting quality dimensions and the second subset of data that does not have conflicting quality dimensions.

[0060] To further explain using a specific task as an example, taking the SFT task as a candidate task, the CAM matrix activates two pairs of quality dimensions: (diversity, consistency) and (accuracy, diversity). Includes (diversity, consistency) and (accuracy, diversity). However, (accuracy, timeliness) is not activated.

[0061] In step S25, the first weight corresponding to each data in the first data subset is determined according to the quality dimension of the conflict.

[0062] The first weight can be calculated based on the marginal contribution information of conflicting quality dimension pairs to ensure that the advantages of high-quality dimensions are preserved while effectively suppressing the negative impact of conflicting dimensions during the optimization process. Alternatively, the first weight can be determined through a weighted fusion method based on the interaction relationship between different quality dimension pairs.

[0063] In some implementations, determining the first weight corresponding to each data item in the first data subset based on the quality dimension of the conflict includes: obtaining the marginal contribution information of each data item in the first data subset under each quality dimension of the conflict based on the quality dimension of the conflict, and determining the marginal contribution information as the first weight of each data item under each quality dimension of the conflict.

[0064] Marginal contribution information refers to the degree to which the performance of a data point on a specific quality dimension contributes to the overall task objective. For example, for a task that requires both high diversity and high consistency, a data point may provide a new sample distribution in the diversity dimension, but may deviate from the labeling standard in the consistency dimension. Its marginal contribution information needs to comprehensively evaluate the balance between these two dimensions.

[0065] Specifically, for each pair of quality dimension pairs with conflicting quality dimensions, the Pareto subproblem for each pair of quality dimension pairs is calculated to obtain the first weight, as shown in the following formula: (3) in, For sample x m The sample weights represent the degree to which a sample is retained. Sample x m Similarly, for the marginal gradient of the i-th quality dimension, Sample x m For the marginal gradient of the j-th quality dimension, T j Let be the lower bound of the constraints for the j-th quality dimension, and st be the constraint identifier for the optimization problem.

[0066] Based on formula (3), the marginal contribution value of dimension i is maximized under the constraint that dimension j does not decrease.

[0067] In step S26, the second weight corresponding to each data point in the second data subset is determined based on the weight of each quality dimension.

[0068] The second weight can be determined by weighted summation of the data in the second data subset. The formula for weighted summation is as follows: (4) in, This refers to the weight coefficient of dimension i, used to characterize the degree of importance attached to this quality dimension. The larger the value, the higher the level of importance attached to it; conversely, the smaller the value, the lower the level of importance attached to it. The weighting coefficients can be preset or user-defined.

[0069] The second weight corresponding to each data point in the second data subset is determined based on formula (4).

[0070] In step S27, the data in the candidate dataset is filtered according to the first weight and the second weight to obtain the target dataset.

[0071] The filtering method can be based on the magnitude of the first weight and the second weight to filter the data, or it can be based on a preset solution objective to solve for the first weight and the second weight, thereby determining the final weight of different data, and then filtering the data based on the ranking of the final weights.

[0072] For example, the objective is to solve for: (5) in, For the Pareto subproblem corresponding to dimension pair (i, j), To activate the normalized weighted coefficients of the conflict pair (i, j), To activate the weighted overall objective of the conflict dimension, as the first weight, The inactive dimensions are jointly maximized as the objective function, and used as the second weight.

[0073] The two parts are jointly optimized using formula (5), and the final weight w is obtained by solving the projective gradient descent. Based on... This is used to filter the candidate datasets.

[0074] In some implementations, the step of filtering the data in the candidate dataset according to the first weight and the second weight to obtain the target dataset includes: determining a target weight for each data item in the candidate dataset according to the first weight and the second weight; sorting the data in the candidate dataset according to the target weight; and determining the target dataset based on the sorted data.

[0075] In this embodiment, the target weight of each data point is determined based on the first weight and the second weight. Then, the data in the candidate dataset is filtered based on the target weight. Different filtering methods are used for different subsets of data during the filtering process, which avoids wasting computing power and meaningless weight trade-offs, and improves the accuracy of data filtering.

[0076] In the specific implementation process, this solution is described as an example of adaptive optimization of the SFT dataset in the financial field. The candidate SFT dataset contains 1 million records, including 60% customer dialogue logs, 30% public financial Q&A, and 10% synthetic data. The target task is the 7B financial vertical model SFT, and the downstream evaluation set is FinanceQA-CN. The conflict information coefficient of the quality dimension pairs under the SFT task conditions is determined, and the threshold is 0.3.

[0077] , ,

[0078] Task profile information Diversity = +1, Consistency = +1, Accuracy = +1, Timeliness = -1, Compliance = +1, Pollution-free = +1.

[0079] Based on conflict information and task profile information, conflict activation information is determined, i.e., it is activated. and Not activated .

[0080] Then, a Pareto trade-off was made between diversity and consistency, with the goal of maximizing the diversity entropy under the constraint that consistency is greater than or equal to 0.85, ultimately retaining 230,000 samples.

[0081] FinanceQA-CN's evaluation metric improved from a baseline of 41.3 to 48.1, while the weighted summation baseline method achieved an evaluation metric of 44.7 for the same data size. (Target dataset after filtering) (230,000 records), and can also output sample-level gradient reports, and mark the marginal contribution of each sample to each dimension.

[0082] In the solution provided in this application, the data screening process is decomposed into constrained optimization on the conflict dimension and joint maximization on the non-conflict dimension by coupling task profile information and conflict information. This ensures that key bottom-line indicators are not breached and maximizes the comprehensive benefits across multiple dimensions, avoiding the pseudo-Pareto problem caused by traditional weighted summation. Furthermore, by introducing an ablation experiment feedback loop, conflict information can be continuously updated with new samples, ensuring that the screening strategy remains adaptive and effective under different tasks and data distributions.

[0083] Figure 3 This is a block diagram illustrating a dataset filtering device according to an exemplary embodiment. Figure 3 The dataset filtering device 300 includes an information acquisition module 310, a data filtering module 320, and a data acquisition module 330.

[0084] The information acquisition module 310 is configured to acquire task profile information of candidate tasks, wherein the task profile information represents the priority of data in at least one quality dimension. The data filtering module 320 is configured to determine the conflict activation information of the candidate task based on the task profile information and conflict information. The data acquisition module 330 is configured to filter the candidate datasets of the candidate tasks based on the conflict activation information to obtain the target dataset.

[0085] Optionally, the data acquisition module 330 is further configured to determine, based on the conflict activation information, a first subset of data in the candidate dataset that has a conflict quality dimension and a second subset of data that does not have a conflict quality dimension. Based on the quality dimension of the conflict, determine the first weight corresponding to each data item in the first data subset; Based on the weight of each quality dimension, determine the second weight corresponding to each data point in the second data subset; The data in the candidate dataset is filtered according to the first weight and the second weight to obtain the target dataset.

[0086] Optionally, the data acquisition module 330 is further configured to obtain marginal contribution information of each data point in the first data subset under each conflict quality dimension based on the conflict quality dimension, and to determine the marginal contribution information as the first weight of each data point in each conflict quality dimension.

[0087] Optionally, the data acquisition module 330 is further configured to determine the target weight of each data item in the candidate dataset based on the first weight and the second weight; The data in the candidate dataset are sorted according to the target weights, and the target dataset is determined based on the sorted data.

[0088] Optionally, the information acquisition module 310 is further configured to acquire task information of the candidate task, the task information including task type and task metadata; Based on the task information and quality dimensions, the task profile information of the candidate tasks is determined.

[0089] Optionally, the dataset filtering device 300 further includes a historical sample determination module, a conflict coefficient acquisition module, and a conflict data filtering module. The historical sample determination module is configured to determine historical sample pairs based on historical datasets, historical quality dimensions, and ablation experiment result sets. The conflict coefficient acquisition module is configured to acquire the conflict coefficient of each historical sample pair for each historical quality dimension. The conflict data filtering module is configured to determine the conflict information based on the conflict coefficient.

[0090] Optionally, the dataset filtering device 300 further includes a result determination module and an information update module. The result determination module is configured to determine the latest ablation experiment result set based on the target dataset; the information update module is configured to update the conflict information based on the latest ablation experiment result set.

[0091] Figure 4 This is a block diagram illustrating an electronic device according to an exemplary embodiment. Figure 4 As shown, the electronic device 400 may include a processor 401 and a memory 402. The electronic device 400 may also include one or more of a multimedia component 403, an input / output (I / O) interface 404, and a communication component 405.

[0092] The processor 401 controls the overall operation of the electronic device 400 to complete all or part of the steps in the dataset filtering method described above. The memory 402 stores various types of data to support the operation of the electronic device 400. This data may include, for example, instructions for any application or method operating on the electronic device 400, and application-related data such as contact data, sent and received messages, pictures, audio, video, etc. The memory 402 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 403 may include a screen and audio components. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in memory 402 or transmitted via communication component 405. The audio component also includes at least one speaker for outputting audio signals. I / O interface 404 provides an interface between processor 401 and other interface modules, such as a keyboard, mouse, buttons, etc. These buttons may be virtual or physical buttons. Communication component 405 is used for wired or wireless communication between the electronic device 400 and other devices. Wireless communication may include Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof; therefore, the corresponding communication component 405 may include a Wi-Fi module, a Bluetooth module, or an NFC module.

[0093] In an exemplary embodiment, the electronic device 400 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the dataset filtering method described above.

[0094] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, which, when executed by a processor, implement the steps of the dataset filtering method described above. For example, the computer-readable storage medium may be the memory 402 including the program instructions described above, which may be executed by the processor 401 of the electronic device 400 to complete the dataset filtering method described above.

[0095] In another exemplary embodiment, a computer program product is also provided, which includes a computer program executable by a processor, which, when executed by the processor, implements the steps of the dataset filtering method described above.

[0096] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0097] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0098] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

Claims

1. A method for filtering a dataset, characterized in that, The method includes: Obtain task profile information for candidate tasks, wherein the task profile information represents the priority of data in at least one quality dimension; Based on the task profile information and conflict information, determine the conflict activation information of the candidate task; Based on the conflict activation information, the candidate datasets for the candidate tasks are filtered to obtain the target dataset.

2. The dataset filtering method according to claim 1, characterized in that, The step of filtering the candidate datasets for the candidate tasks based on the conflict activation information to obtain the target dataset includes: Based on the conflict activation information, a first subset of data in the candidate dataset that has a conflict quality dimension and a second subset of data that does not have a conflict quality dimension are determined. Based on the quality dimension of the conflict, determine the first weight corresponding to each data item in the first data subset; Based on the weight of each quality dimension, determine the second weight corresponding to each data point in the second data subset; The data in the candidate dataset is filtered according to the first weight and the second weight to obtain the target dataset.

3. The dataset filtering method according to claim 2, characterized in that, The step of determining the first weight corresponding to each data item in the first data subset based on the quality dimension of the conflict includes: Based on the quality dimension of the conflict, obtain the marginal contribution information of each data point in the first data subset under the quality dimension of each conflict, and determine the marginal contribution information as the first weight of each data point in the quality dimension of each conflict.

4. The dataset filtering method according to claim 2, characterized in that, The step of filtering the data in the candidate dataset according to the first weight and the second weight to obtain the target dataset includes: Based on the first weight and the second weight, determine the target weight for each data item in the candidate dataset; The data in the candidate dataset are sorted according to the target weights, and the target dataset is determined based on the sorted data.

5. The method for filtering datasets according to any one of claims 1-4, characterized in that, The process of obtaining the task profile information of the candidate tasks includes: Obtain the task information of the candidate task, the task information including task type and task metadata; Based on the task information and quality dimensions, the task profile information of the candidate tasks is determined.

6. The method for filtering datasets according to any one of claims 1-4, characterized in that, The method further includes: Historical sample pairs were determined based on historical datasets, historical quality dimensions, and ablation experiment result sets; Obtain the conflict coefficient for each pair of historical samples for each pair of historical quality dimensions; The conflict information is determined based on the conflict coefficient.

7. The dataset filtering method according to claim 6, characterized in that, The method further includes: Based on the target dataset, determine the latest ablation experiment result set; The conflict information is updated based on the latest ablation experiment results set.

8. A dataset filtering device, characterized in that, The device includes: The information acquisition module is configured to acquire task profile information of candidate tasks, wherein the task profile information represents the priority of data in at least one quality dimension. The data filtering module is configured to determine the conflict activation information of the candidate tasks based on the task profile information and conflict information. The data acquisition module is configured to filter the candidate datasets of the candidate tasks based on the conflict activation information to obtain the target dataset.

9. An electronic device, characterized in that, include: A memory on which computer programs are stored; A processor for executing the computer program in the memory to implement the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method described in any one of claims 1-7.

11. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1-7.

Images